Assessment of Significant Change in BMD: A New Approach

We wish to comment on the paper by Ravaud et al. on the assessment of “minimal detectable” significant change of bone mineral density (BMD) for an individual.1 Our data and other data suggest, contrary to their assertion, that measurement error is related to the actual value, increasing with lower BMD. Thus, a universal cut-off value for the diagnosis of change for all individuals is unlikely to be informative. Moreover, the assessment of change for an individual should take into account the serial correlation between measurements within an individual. These concepts lead us to propose a new formulation of minimal detectable difference that is more realistic and clinically useful in assessing progress of osteoporosis and response to therapy. In the study of Ravaud et al., the reproducibility of BMD measurements in elderly postmenopausal women (average age 79.5 years) was poorer than that in younger postmenopausal women (average age 52.8 years). However, they concluded that the error of measurement was independent of BMD level and hence suggested that the determination of cut-off values for the minimal detectable difference and precision should be based on absolute units rather than percentage. We have shown that the SEM of femoral neck and lumbar spine BMD in women aged 70 years or older (mean age 75 years) was approximately twice that in younger women (mean age 46 years).2 Furthermore, in that study, the absolute within-subject difference in femoral neck BMD was negatively related to their mean (r = −0.4; p < 0.001), indicating that the measurement error was higher among those with lower BMD. We and others also have observed that the rate of change in BMD is inversely correlated with its initial value, such that subjects with higher baseline BMD have higher rates of loss, consistent with the phenomenon of regression-toward-the-mean.3-5 Therefore, the diagnosis of change in BMD for an individual cannot be based on a universal cut-off value. The above formulation appears to be too conservative, based on a simulation of 100,000 subjects with mean femoral neck BMD of 0.78, SD of 0.15 g/cm2, and SEM of 0.012 g/cm2. With a “true” loss of 0.03 g/cm2 in all subjects in the simulated population, the LSC formula identified correctly only 39% of subjects as having “significant bone loss,” creating an error of 61%. According to a recent systematic review of efficacy of alendronate clinical trials, among patients who were treated with alendronate, the typical increase in lumbar spine and femoral neck BMD varied between 1% and 3.5% (or 0.01–0.028 g/cm2).6 Similarly, in the Dubbo study, elderly women aged 60+ years lost an average of 0.018 g/cm2 femoral neck BMD over an average of 2-year follow-up, and only 36% of the women lost more than 0.033 g/cm2 during the period. Clearly, the LSC criterion is too strict and its application can lead to incorrect diagnosis of change. Thus, the LSC can be calculated as For an SEM of 0.012 g/cm2 and a serial correlation of 0.88, the LSC is 0.012 g/cm2 or 0.021 g/cm2 lower than that under the assumption of no serial correlation. This new formulation correctly identified 81% of subjects as having significant bone loss in the simulation and 50% in the Dubbo study. Ravaud et al. address an important issue; however, it appears that none of the CVs reported in Table 2 is consistent with the calculation described on page 1450, given the means reported in Table 1 and the SDs in Table 2. In fact, all CVs reported are lower than they should be. It is mathematically incorrect to calculate the mean of SDs or CVs as they describe (page 1452), because these are square-root quantities and they are not amenable to simple addition. This, although often found in the osteoporosis literature, leads to an underestimated CV.9-12 Also, the confidence interval for the SD of lumbar spine BMD in “healthy postmenopausal women” would appear to be 0.007–0.01 g/cm2 and not 0.00412–0.00576 g/cm2 as they report. In summary, the diagnosis of change based on the traditional LSC formulation is likely to overestimate the influence of random variability and miss clinically relevant and statistically significant changes for an individual. We propose that the formulation take into account the serial correlation between measurements within an individual resulting in considerably narrower criteria for such changes.